Corepresentational Grammar and Parsing English Comparatives Karen P#an University of )linnesota SEC. 1 INTRODUCTION SEC. 3 COREPRESENTATIONAL GRAMMAR (CORG) Marcus [3] notes that the syntax of English comparative constructions is highly complex, and claims that both syntactic end semantic information must be available for them to be parsed. This paper argues that comparatives can be structurally analyzed on the basis of syntactic information alone via a strictly surface-based grammar. Such a grammar is given in Ryan [5], based on the co- representational model of Kac Ill. While the grammar does not define a parsing algorithm per se, it nonethe- less expresses regularities of surface organization and its relationship to semantic interpretation that an ade- quate parser would be expected to incorporate. This paper will discuss four problem areas in the description of comparatives and will outline the sections of the grammar of [5] that apply to them. The central problem in parsing comparatives involves identifying the arguments of comparative predicates, and the relations borne by these arguments to such predi- cates. A corepresentational grammar is explicitly de- signed to assign predicate-argument structure to sen- tences on the basis of their surface syntactic organi- zation. SEC. 2 COMPARATIVE PREDICATES An initial assumption underlying the proposed analysis of, comparatives is that the comparative elements such as ~r~' faster, more spacious, are syntactically akin to icat-~, and thus that the principles applying to predicate-argument structure extend to them. Each com- parative element will accordingly have arguments (Subject and Object) assigned to it, and comparative predications will also be analyzed as being in relations of subordin- ation or superordination with other predications in the sentences in which they appear. For example, in (l) below, the comparative predicate richer will have both a simple NP Subject and a simple NP~t: (1) John knows doctors richer than Tom SUBJ ~" OBJ The referent of OBJ(richer), i.e. Tom, is to be inter- preted as the standar d-o-~-compariso-n-against which the referen~ of doctors is Judged. The entire predication forms a term ~ion ('T') acting as OBJ(kn~ow), so that the whole relational analysis is as shown In (2). (2) John knows doctors richer than Tom I T suBJ T 0~J Pr/richer(T} su~J 0~J Because Pr/richer is included in an argument of another predicate (~ the former is in a relation subordinate to the latter. This analysis assumes three types of comparative predi- cates: adverbial, adjectival, and quantifier. Illustra- tions are given below: (3) Alice builds planes faster than robots fly them (4) John met people taller than Bob (5) Alice drank more beer than Helen The adverbial predicates are subcategorized as taking predicational arguments in both relations, and only such arguments; the other types can take nonpredicational arguments, though in some cases their Objects may be predicational. The grammar itself consists of two sets of principles. The first set consists of general constraints on sentence structure and applies as well to non-comparative con- structions. These principles are discussed in detail in [l] and [2] and will be presented here without justifi- cation. In addition there are a number of principles applying only to comparative constructions but non ad hoc in the sense that each can be applied toward the so- lution of a number of distinct problems of analysis. These principles are as follows: (6) Law of Correspondence Every NP or term in a sentence must be assigned a relational role. Ill (7) L~wof Uniqueness No two elements in a sentence may bear the same relation to a sinnle predicate unless they are coordinate or coreferential. Ill (8) Object Rule (OR) If P is an active transitive predicate~ OBJ(P) must be identified in such a way as to guarantee that as many segments thereof as possible occur to the right of P. Ill (g) ?~ulti-Predicate Constraint Every predicate in a sentence which contains more than one predicate must be in an ordination relation with some other predicate in that sentence.[4] (lO) Term Identification Principles a. Any predication with the internal structure OBJ-SUB-PREO may be analyzed as T. Any UP is a T. Any T satisfying either of these conditions is a SIMPLE TE~I. b. Any predication consisting solely of a compara- tive predicate with simple ~!P's as arguments is a T; such expressions will be called SIMPLE CO?IPARATIVE TE~.IS. All others will be COtlPLEX COMPARATIVE TE~IS. c. Any predication whose Subject occurs to the right of than, and whose predicate either occurs tot E~-e left of than or occurs as SUBJ(do) where do itself occursto the right of than, is a T; s~h expressions will be called PRE-'DTCATE- CONTAIN~IG TERMS or PCT's. (ll) Comparative Object Rule The object of a comparative predicate is any term or predication satisfying the subcategorization of the predicate and which in- cludes some element occurin 0 immediately to the right of than. (12) Comparative-e-~ubject Rule The Subject of a compara- tive predicate must occur to the left of than. (13) Comparative Object Restriction The Object o-? a nonadverbial comparative predicate must be a simple term unless the tiP occuring immediately to the right of than is SUBJ of a PCT; in that case, the OBJ of the non-adverbial comparative predicate must be a PC-term. These principles do not define a parsing algorithm per se; rather, they express certain surface true restric- tions which taken together and in concert with the gen- eral principles from Kac Zl ] and [2 ], define exactly the set of predicate argument structures assignable to a comparative construction. Since no particular analyt- ic procedure is associated with CORG, the assignment of particular analyses may be thought of either as a com- parison of complete potential relational analyses with the principles, whereby all potential analyses of the string not consistent with the grammar are discarded, or as a process of sequential assignments of partial analy- ses where each step is checked against the principles. The sequential method of analysis will be used here to present the operation of these principles; however, it is not a necessary adjunct to the grammar. 13 SEC. 4.0 STRUCTURE TYPES AND DESCRIPTIVE PROBLEMS There are three types of comparative predicates, already noted in section 2: adjectival, quantifier and adverbial. The differing subcategorization of these predicates does affect the possible analyses for a given sentence. Sev- eral other factors which influence the interpretation of the sentence are the position of the comparative predi- cate in the sentence, the degree of ellipsis in the than-phrase, and the subcategorization of surrounding p-~-~dicates. The effect of the type of predicate and the effect of the position of the predicate (in particular relative to than) will be considered separately in the following sect~o "-ns. The effects of the degree of ellipsis in the ~than phrase and the subcategorization of surrounding predlcates will be considered together in section 4.3. It should be kept in mind however that all of these variables may act together in any combination to affect the type and number of interpretations a given sentence may have. SEC. 4.I SUBCATEGORI~.ATION AND PREDICATE TYPES The. effects of the type of comparative predicate on the interpretation can be noted in (3) and (4). The adverb- ial predicate faster in (3) takes predicational arguments only (ignoring f-T6"r"now the problem of lexical ambiguity) while the adjectival predicate taller takes non-predica- tional (.gP or Term) arguments. To see how these differences interact with the possible analyses which may be assigned, consider a complete analysis of (4). This analysis may begin with any ele- ment in the sentence. In most cases the assignment of the object of the comparative predicate, as the first step, will result in a more direct path to a complete analysis. Assume then, that Bob has been analyzed as O~(taller). This assignment-~atisfies the Comparative ObjecT~-uTe and is also consistent with the OR. (14) John met people taller than Bob. T Since neither met nor taller is a reflexive predicate, the Law of Unique' 'ness guarantees that Bob cannot be analyzed as OBJ (P), where P is any pr~-'Tcate (other than taller) as long as it is analyzed as OBJ(taller). Slnce t-TEe'F'~ are two non-reflexive predicates in this sentence (taller and m e_~.t), there are four remaininq re- lational ass-~g~ents whlch must be made before the analy- sis is complete. These are SUBJ(me_~.t), OBJ(met), SUBJ (taller) and some ordination relatlon betwee n-the pred- icates met and taller. John or Either ~ people may be analyzed as SUBJ(taller) at this point since both satisfy the Comparative ~-~t Rule by occuring to the left of than. If John were assigned the relation SUBJ(taller-) The analysis would violate some principles. A~for purposes of demon- stration, that John=SUBJ(taller). The relational analy- sis at this point would th en be: (15) John met people taller than Bob SRBJ T o~J The remaining relational assignments would be OBJ(met), SUBJ(met) and some ordination relation for the two pred- icate~ The next apparently logical step would be to analyze people as O~j(me_~t). However, this will violate the OR, since it is possible to include mere than just the ;(P people as part of the OBJ(met). The OR requires that as many segments as possible-Eccuring to the rioht of a predicate be included in OBJ(P). The way to satis- fy this condition would be to analyze ~ as part of PR/taller. Then the OR would be satisfied by the maxi- mum number of elements (consistent with the grammar) which occur to the right of met. The only possible re- lation that people could bear to taller would be SUBJ (taller) sin~occurs to the l~ than (see Com- parative Subject Rule). If it is analyzed as SUBJ(tal- • ler), then John can no longer be analyzed as SUBJ(talL ler). These steps would wive the following partial rela- tional representation: (16) John met people taller than Bob T SUBJ ~ OBJ PR/taller(T) OBj At this point in the analysis, the only relation which needs to be assigned still is SUBJ(met). The assignment of this relation to John is the only possible choice which violates no principle of the grammar and this as- signment would give a complete analysis. The analysis of (3) procedes along somewhat different lines due to the subcategorization of the adverbial comparative predicate faster, which requires predica- tional arguments. Thean~sis can begin as before by attempting to assign arguments to the comparative predi- cate faster. However, the first NP after than cannot be assigned to faster as OBJ since it is not a predicational arnument. The subcategorization of faster requires com- plete predications to be available b~arguments for it may be identified. Thus consider the other predi- cates, build and fly. Both are transitive predicates taking on ~simple HP's as arguments. The ~IP them must be analyzed as OBJ(fly) because of the OR. Th~mpar- ative OBJ Rule and ~ OR together will require robots to be analyzed as part of the PR/fly. Since robots occurs immediately to the right of than, it mus-Et-6"~in- cluded as part of the OBJ(faster) by ~Te Comparative OBJ Rule. The OR requires the"O-~J-~f any predicate to in- clude as many elements to the right of that predicate as possible. Therefore, if possible, fly and them must also be included as elements of OBJ~-?aster) ~ince faster is an adverbial predicate, itwl-'~TTT-allow a com- pe-l-eEe-predication (in fact requires) to be its object. Thus, all three of these aspects of the grammar work to- gether to force the string robots fly them to be anal- yzed as a predication PR/fly as shown below, with PR/fly analyzed as OBJ(faster)(as allowed by the Comparative OBJ Rule). (17) Alice builds planes faster than robots fly them T SUBJ OBj I" PR/flv OBJ At this point the arguments of build still need to be assigned and build and faster must be assigned some or- dination rela~ Sln~ter requires a complete predication for its subjec~ predication build must be built first. If any rip's other than AliceTplanes are used as arguments for builds, the anay T'~s cou~ be completed. For example~obots were analyzed as OBJ(bullds) (as well as SUBJ(fly-]~-T, then either Alice or SlCOUld be analyzed as SUBJ(builds) completing d. (18) Alice builds planes faster than rgbo~s fly them SU~J I" "F OBa S~BJq" ~Bj PR/build PR/fly OBj PR/build could then be analyzed as SUBJ(faster) and all the necessary relations between arguments and predicates, and between predicates themselves(i.e, ordination rela- tions) would be assigned. However, the analysis would be ill-formed since one element, in this case lap_~, would be left unanalyzed in violation of the Law o? ~orrespon- dence. The only way this situation can be avoided, while at the same time not violating the OR or the Comparative Object Rule as discussed above for the OBJ(faster), would be to use only Alice and planes as arguments for builds. The OR would requlr~ that~.~ be analyzed as OB~ ~ (builds) leaving Alice to be analyzed as SUBJ(builds). This resulting pred dT~'ation Pr/builds can then be anal- yzed as SUBJ(faster) completing the analysis with all rules in the grammar satisfied. (Ig) Alice bu~ds planes faster than robots fly them SU~V T OBj ~ SHR,/ "r' onj PR/builds SUBJ I P~/fIY OBJ 14 The most obvious differences between the analyses of (3) and (4) is in the types of arguments which the compara- tive predicates take and the ordination relations be- tween the predicates and the order in which the differ- ent predications were "built up". For (3), the argu- ments for the non-comparative predicates must be assigned first, before the arguments for the comparative predi- cate. This is required by the subcategorization of the adverbial predicate, which takes predicational arguments only. In this sentence, the non-comparative predicates are analyzed as subordinate to the comparative predicate. This too is a conseqence of the subcategorization of faster. For (4), the most efficient procedure for as ~ing relations (i.e. the one requiring the least backtracking) requires the arguments of the comparative predicate taller to be assigned first. In addition since the~egorization of this predicate allows only for non-predicational arguments, the comparative predicate is analyzed as subordinate to the non-compar- ative predicate in the sentence. Thus the type of com- parative predicate and its subcategorization affects the type of analysis provided by the grammar, and also the "optimal" order of relational assignments, when proce- dural aspects of the analysis are considered. SEC. 4.2 POSITION OF THE COMPARATIVE PREDICATE There are two aspects to the problem of the position of the comparative predicate: one involves the position of the SUBJ(COMP P) relative to than; the other involves the position of the entire comparative predication rela- tive to any other predicate in the string. SEC. 4.2.1 COORDI~IATE AND NON-COORDINATE ADVERBIAL COMPARATIVE CONSTRUCTIONS In some cases, the arguments of comparative predicates may be coordinate. This will always be the case for adverbial comparative predicates for which there is some ellipsis in the string as in (20) John builds planes faster than robots Here robots can be considered to be coordinate with either E-'~es or John, that is it can be interpreted as either t h-e~-O'BJ(b~s) or as the OBd(builds). In non- adverbial comparative constructions, it will not always be the case that a single riP after than will be inter- preted as coordinate with some nother-"r-~TP. Consider the differences in possible interpretations between (4) and (21) (21) John met taller people than Bob (4) John met people taller than Bob For (4), there is only one possible interpretation, while there are two possible interpretations for (21). That is, in (21) Bob may simply be interpreted as OBJ(taller) correspond dTng to the meaning of the sentence (22) John met people who are taller than Bob However, (21) has another interpretation in which Bob is interpreted as SUBJ(met). This case corresponds t~he interpretation of (23). (23) John met taller people than Bob did For this second interpretation, there are two subjects for me.__tt, i.e., John and Bob. This means that John and Bob must be forma aITy def~d as coordinate arguments. l~-~'s formal definition is necessary since the Law of Uniqueness states that no two NP's may bear the same relation to a predicate (i.e. both be SUBJ(P i) unless they are coordinate or coreferentia1. Such a definition for rlP's such as John and Bob in (23) is not unreason- able since they bo Eh meet ~ basic requirements for coordinate elements. They are both interpretable as bearing the same relation to some Predicate Pi. The Comparative Object Restriction and a definition of coordinate comparative elements are required to precise- ly define the conditions under which two elements may be construed as coordinate in a comparative construction. The essence of the Coordinate Comparative Definition (not included here due to space considerations) is that any two elements may be coordinated by than if no non-adverbial comparative predicate occurs immediately to the left of than. The ultimate consequence of this condition is that only one interpretation is a11owed for constructions like (4) and this interpretation does not include any arguments coordinated by than. This means that in (4) for example there is no possl-' %le analysis in which Bob can be SUBJ(met). In the coordinate interpretation of (22), (i.e., where John is coordinate with Bob) the final analysis of the s-ErTng will include the ~r6Tlowing predicational struc- ture: (24) John ~t taller pe?pleOBJ thans~ Pr/met(PCT) It is this term, then, which is assigned to the relation OBJ(taller), ~ being SUBJ(taller) (note that people plays two distlnct roles in this sentence). (25) John met taller peopl~ than Bqb "I ~ ~ OBQ SOBJ • F" " Pr/met(PCT) L SUBJ OBJ This particular assignment (of pr/met as OBJ(taller~ is allowed by the Comparative Object Restriction. That is, taller, being non-adverbial comparative predicate, is ~bcategorized for predicational arguments. But in (25) OBJ(taller) contains a predicate as one of its arguments. This particular predicational structure is defined as a Predicate Containing Term or PCT by the Term Definition~ The Comparative Object Restriction has the effect of al- lowing the OBJ(CO~P P) to be a PCT. Since the particular substring of (22), met people Bob need not be analyzed as a PCT, an altern~ive analysis for (22) is also pos- sible. The alternative analysis would be like that for (4), where only Beb=SUBJ(taller). That is, the Compar- ative Object Restriction does not necessarily require an analysis for (22) like (25); it merely allows it if cer- tai:n conditions set out in the Term Definition are met. The Comparative Object Restriction is quite important, then, in distinguishing the possible analysis for non- adverbial comparative constructions. It is equally Im- plant in obtaining the correct analysis for the sen- tence types to be discussed in the next section. SEC. 4.2.2 SUBJECT COMPARATIVES The position of the entire comparative predication, rela- tive to other predicates in the string is also quite im- portant in determining the possible types of analysis. Sentence (25) exhibits a subject comparative where the comparative predication occurs to the left of another predicate. It is useful to compare this sentence with the object comparative in (22) repeated here. (26) Taller people than Bob met John (22) John n~t taller people than Bob As has already been discussed in 4.2.1, (22) has two pos- sible interpretations. Sentence (26), however, has only one possible interpretation. Therefore there should be only one possible analysis. The analysis which needs to be avoided is (27) Taller people thans~ ~ m~ John T o~J I pr/m@t SUBJ OBJ This case must be disallowed while at the same time al- lowing the structure in (24) to be analyzed as OBJ(tal- ler). The Comparative Object Rule and the Term 15 Definitions work together to achieve this. The structure Pr/met shown in (28) does not meet the requirements set out for a PC-Term and the subcategorization of taller (i.e. non-predicational arguments only) will not allow Pr/met to be analyzed as an argument of taller unless it is analyzable as a PC-Term. Thus, the subcategorization of taller and the Comparative Object Restriction will both prevent the assignment of Pr/met as OBJ(taller)in (27). Since an analysis which includes (27) is not pos- sible, the only way the analysis can procede is as fol- lows. The Comparative Subject Rule will require people=SUBJ(taller) since it is the only tip to the left of than. Since Bob is the element occuring immediately to t-'h-e-right of~n, it is the only ~IP which can be analyzed as objec-'t ~f taller. The resulting predication Pr/taller is defined as a term by (IOb). (28) Taller peqple than B b met John ¢ s J Pr/taller(T) The MP John must be analyzed as OBJ(met) to satisfy the OR, leav-~Pr/taller to be analyzed as SUBJ(met). This will also satisfy the )lultiPredicate Constraint since taller and met will be in some ordlnatlon relation as a res-'~. (2g) TallerLprxtaller(T)su)dpeqple~uB,] than ~jB b m it JofnOBd Pr/met No other analysis is possible since no non-comparative predicate occurs to the left of than (which would allow for possible coordinate interpretatl ~ons). SEC. 4.2.3 COMCLUSIONS The important points in this section are that for Sub- ject Comparatives such as (26), only one interpretation is possible, while for Object Comparatives such as {21), two interpretations are possible. Position of the com- parative predication relative to the rest of the string is thus an important factor in determining the number of possible interpretations. Position of individual NP's relative to than is also an important factor in deter- mining the number of possible interpretations a sentence may have; Sentences like (4),where no tIP occurs between than and the comparative predicate, have only one inter- pretation, ~lhile sentences like (ZIP, where an PIP does occur in the position, have two possible interpretations. The Comparative Object Restriction and the Term Defini- tions figure crucially in all these cases in the deter- mination of the correct number and type of possible analyses. SEC. 4.3 DEGREE OF ELLIPSIS AND SUBCATEGORIZATION O.~F SURROUtlDIr~G PREDICATES The degree of ellipsis following than in comparative structures is quite important in ~rmining the number of possible interpretations a structure may have. For example, in the first sentence of each pair below, where only a single predicate occurs before than, more than one interpretation is possible per str-~, while in the second sentence in each pair, where an PIP followed by some predicate occurs, only one interpretation is possible. (30) Alice builds planes faster than robots (31) Alice builds planes faster than robots do (32) John knows richer doctors than Alice (33) John knows richer doctors than Alice does The actual analysis of these sentences will not be presented here. Such sentences are discussed in detail in Ryan [5]. SEC. 4.3.1 DEGREE OF ELLIPSIS AND SUBCATEGORIZATION OF SURRDUMDING PREDICATES. The problem of degree of ellipsis interacts crucially with another factor, the subcateqorization of surround- ing predicates, in a very interesting way. Consider , the following sets of sentences. (34) John knows more doctors than lawyers debate (35) John knows more doctors than lawyer s debate psychiatrists (36) John knows more doctors than lawyersrun (37) John knows more doctors than lawyers spoke to (38) John hired more doctors than lawyers debate (39) *John hired more doctorsthan lawyers debate psychiatrists (40) *John hired more doctors than lawyers run (41) John hired more doctors than lawyers spoke to (42) John thinks more doctors than lawyers debate (43) John thinks more doctors than lawyers debate psychiatrists (44) John thinks more doctors than lawyers run (45) *John thinks more doctors than lawyers spoke to These sentences contain different combinations of com- parative predicates with either transitive or intrans- itive verbs following them and preceding verbs which take: either complement or NP objects (34~-(37); NP objects only (38-41); and complement objects only (42- 45). The type and number of interpretations depends on the subcategorlzation of these verbs and the verbs fol- lowing the comparative predicate. The flrst sentence in each group contains a transitive verb, debate, with no overt object. The second sentence in eac~group contains debate with an overt object. This results in (39) in an ungrammatical sentence, as compared with (38), and in (35) in a sentence with only one possible interpretation as compared with (34), which has two possible interpre- tations. The third sentence in each group contains an intransitive verb, run. This also results in an ungram- matical sentence for T40) in the second group and in a sentence with only one interpretation, (36) in the first group. The last sentence in each group contains another transitive verb, spoke to, without an overt object. The difference between this~erb and debate is that debate is a so-called 'object deletable've-~'eF~-while spo]E~"~o- is not. Mote that in (45) this results in an ungra~at- lcal sentence (compare to 42) while in (37) the sentence is grammatical. However, in (37) the structure of the phrase more doctors than lawyers differs from its struc- ture in (35) and (36), in which more doctors than ~e tS the subject of the third verb. That is not in (37), where only la~ers is the subject of the third verb. It can be seen from this that the sub- categorization of the preceding the following predicates Is very Inq~ortant to the structure of the comparative predication. In addltlo~as the first two sentences in each group show, the degree of ellipsis also affects the structure. In all cases, the structure of the phrase more doctors than lawyers shifts in structure. The most important aspect of this data is the type of arguments which the comparative predicates must take. In these particular cases it is a change in the object of the comparative predicate which corresponds to a shift in the structure of the sentence. This is accounted for most directly by the rules in (lOp, (ll) and (13). For example, in (36) the OBJ(more) is lawyers and the co~q}lete predication Pr/more ~he Surf run. This partial analysis is~wn in (46). (46) John knows more doqtors than lawxers r4n suBJ o~j T Pr/more(T) SUBJ 16 i In (38), the object of more is the sequence doctors lawyers debate, a term according to (lOa). shown in the partial analysis in (47). (47) John hired more doctprs than lawyers debate T )OBJ SUBJ I" | Pr/debate(T) SqBJ ~qj Sentence (36) could not be analyzed as in (47) because run, the third verb in (36), is intransitive while de-e~ate, the third verb in (38), is transitive. Thus run cannot be included in any structure satisfying the Te~ Identification Principles (lO), while debate can be so analyze@. This means that run cannot be T~cluded as part of the OBJ(more). This is ~ranteed by the Comparative Object Restrlct-' ion (13). Both of the analyses shown in (46) and (47) are possible for sentence (34) since knows may take predicational objects (in this case, more doctors than lawyers run) or it may take nonpredicatlonal objects such as the Complex comparative term in (47). Sentences (39) and (40) do not have possible analyses since hired cannot take predicational objects (such as that sho o-wn-in (46)), and the presence of either an intransitive verb (run) or a transitive verb with an overt object (debate'-psychiatrists) after the compara- tive predicate, forces such a structure because of rules (lO) and (13). Sentence (41) would have a structure similar to (47). Sentences (42) - (44) v~uld all have structures similar to the partial analysis in (46). This is forced by the subcategorization of thinks, which takes only predica- tional objects. There iT-no possible analysis for (45) since the subcateqorization of s o_~to, unlike debate, requires the presence of an overt object. But i a?-a-n object is assigned to spoke to, the result will ulti- mately be a structure Ti-Ee'-tlTat shown in (47). But the structure shown in (47) is a term and therefore nonpred- icational. This means it could not be analyzed as OBJ(thinks), while requires a predicational (complement) structure. Finally, it is precisely because a sentence with sooke to as the third verb must have a structure like (~TF TT.e. nonpredicational) that sentence (41) has a possible analysis in contrast to (45). That is, the structure of the string more doctors than lawyers spoke to in (49) has a nonpredicational (comparative term) structure. Since it is a term and not a predication, any verb tak- ing it as an argument must be subcategorized for nonpred- icational arguments. Think in (45) takes only predica- tional arguments in the ~ect relation, while hired in (41) takes only nonpredicational arguments in th-'-e-~'6ject relation. Thus, only the sentence with hired may take the comparative term as an argument. But sooke to does not allow the string more doctors than lawyers to simply be analyzed as its sub-ject, since no possible object would then be available for spoke to, However, if the string more doctors than lawyers is not analyzed as SUBJ(spoke to), it will not be possible to analyze the string as a predication Pr/spoke to, thus blocking the analysis of the string as OBJ(think). SEC. 4.3.2 CONCLUSION The degree of ellipsis and the subcategorization of the surrounding predicates interact to affect the possible number and type of interpretations for each of the sen- tences in this section. That interaction can be most clearly seen in a comparison of (34) and (35) and (36). The verb know is subcategorized for either predicational or nonpred-i-E~tional arguments. This allows the string more doctors than lawyers debate to have two possible structures corresponding to the structures shown in (46) and (47). The.structure in (46) is a predicational structure while the structure in (47) is a nonpredica- tional structure. The subcategorization of knows allows either of those as possible interpretations of the OBJ (knows). Verbs subcategorized for only one type of ar- gument, say predicational, will allow only one of those possible structures of more doctors than lawyers .debate, in this case the predica'tional one shown in (46), to be analyzed as the object of that verb. This is one way in which the subcategorization of surrounding predicates affects the type and number of possible interpretations a sentence may have. The effect of the subcategorization of the following predicate parallels the effect of no ellipsis after than. Thus sentences (36) and (36) each have only one possib bT~ interpretation and the relation of the string more doc- tors than lawyers is the same in each case; that is, it is the same as the predicational structure shown in (46), being the subject of the following predicate. Thus, the presence of an intransitive verb or the presence of a transitive verb plus an overt object to its right as in (35) and (36) forces a predicational structure of the type shown in (46). Since knows takes predicational objects, these sentences are still grammatical. If hired is substituted for knows . as in (39) and (40), the sentences are no longer grammatical, since the subcate- gorization of hired does not allow predication argument~ The last type of effect of the predicate following than is in some cases to force a nonpredicational structure like that shown in (47). The verb s~oke to is not an object deletable verb, while the verb debate does allow unspecified objects. For this reason,~erb sooke to cannot be part of a structure like that shown in-~6), since it would require the object of spoke to to be analyzed as "unspecified". Thus, the presence of a verb like spoke to after than forces the nonpredicational structure o?-the type s-hown in (47), since in this struc- ture the object of ~ to would be overt. Since the presence of spoke to force's a nonpredicational structure for the string more doctors than lawyers spoke to, it can only occur as part of an object of a verb which al- lows nonpredicational objects, like know or hired. It follows from this that if the string more doctors than lawyers spoke to occured after a verb which took predicationa'l arguments only, such as thinks, the result would be an ungrammatical sentence. This is in fact the case, as can be seen from sentence (45). SEC. 5 CONCLUSIONS The rules presented here provide an axiom system which allows only one possible analysis for each interpreta- tion of a sentence, and no possible analysis for sen- tences which are ungrammatical. The rules specifically proposed for comparatives have been shown to apply to a wide variety of construction types; for example, the Comparative Object Restriction and the Term Definitions figure crucially in the analysis of sentences in all the subsections of section 4. In addition, these rules are based on observations about characteristics of the sen- tences which are either directly observable in the string (e.g. left to right relative order) or which are a necessary ~art of any grammatical description (e.g. subclassification and subcategorization of verbs). Such a grammar can provide useful and accessible information for the problem of parsing as well as grammatical description. 17 REFERENCES I. Kac, Michael (1978) Corepr~sentation of Grammatical Structure. Hpls: Uni~rsity of Hlnnesota Press. 2. , (1980) "Corep~sentatlonal Grammar". In Syntax & Semantics 13, E. A. Moravcsik & J. R. Wirth (eds.). Academic Press. 3. Marcus, Mitchell (1980) A Theory of Syntactic Recognitio~ for Natural Languaqe. Cambridge, MA: ~T Press. 4. Rtndflesch, Tom (1978) "The General Structure of Hulti-Predlcatlonal Sentences in Engllsh" in Mlnnesota Papers 5, G. A. Sanders and )l. 8. Kac, eds. 5. Ryan, Karen L. (1981) A Surface Based.Analysis of En91tsh Comparative Constructions. H.A. Thesis, University of Minnesota. 18 . Corepresentational Grammar and Parsing English Comparatives Karen P#an University of )linnesota SEC. 1 INTRODUCTION. SEC. 3 COREPRESENTATIONAL GRAMMAR (CORG) Marcus [3] notes that the syntax of English comparative constructions is highly complex, and claims that both syntactic